Deep learning-based perfusion quantification and large vessel exclusion for renal multi-TI arterial spin labelling...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Renal Physiology

Background:

• Arterial Spin Labeling (ASL) is crucial for renal perfusion assessment.
• Traditional ASL quantification methods struggle with low signal-to-noise ratio and vessel contamination.

Purpose of the Study:

• To develop and validate a deep learning (DL) approach for improved renal perfusion, bolus arrival time (BAT), and bolus length (BL) quantification.
• To exclude large vessels and enhance accuracy in ASL imaging.

Main Methods:

• A BiLSTM-based deep learning network was trained on simulated ASL data.
• The DL model was tested on both simulated and in vivo data, comparing results with traditional Buxton model fitting.
• Manual masks were used for comparison in traditional methods.

Main Results:

• The DL approach demonstrated superior robustness against noise in in vivo data compared to traditional methods.
• DL-based quantification showed less deviation from reference values with fewer data averages.
• DL masks effectively excluded high-perfusion pixels, improving accuracy.

Conclusions:

• The proposed BiLSTM DL network offers a more accurate and noise-robust tool for ASL-based renal perfusion quantification.
• This method shows potential for overcoming limitations of traditional ASL quantification techniques.
• Further investigation is needed to reconcile simulation and in vivo data discrepancies.